The National Energy Research Scientific Computing Center (NERSC) is a computing facility for the U.S. DOE.
Franklin
Franklin is the biggest machine in NERSC. It consists of 9,660 quad-core nodes (38,640 processors). To execute a g-rsm experiment in frankin, following processes are needed.
- Make a symbolic link of gunzip for uncompress command. (for example, $ ln -s /usr/bin/gunzip ~/bin/uncompress)
- Choose "nersc" as institution in the usual installation process. (machine=linux, march=mpi will be chosen)
- Specify "ftn" for XXX_F77 in options-linux-mpi in both libs/opt-libs/ and gsm(rsm)/opt.
- Add "-Msave" for FORT_FLAGS and LOAD_FLAGS in gsm(rsm)/opt/options-linux-mpi.
- Specify nothing for EXTRA_LIBS_XXX in gsm(rsm)/opt/options-linux-mpi.
- Change status(12) to status(8) in libs/lib/w3lib_linux/nainit.f.
Then, do all compilation again (from configure-libs to configure-scr). "qsub" is used for submitting a g-rsm script.
Benchmark
288x350 10km CaRD10v2 experiment for 6 hours.
NPES
|
Wall time
|
ratio
|
512
|
174 sec
|
0.32
|
256
|
288 sec
|
0.53
|
128
|
542 sec
|
1.0
|
Hi, Jan
After reading the SCOAR job scripts. The SCOAR works in this way, the ROMS passes the SST to the coupler. The SST is interpolated to RSM grid, then is converted to grib format. I guess the problem is not in RSM part. But it is possible in the ROMS part or sst_to_grib.
1). Please check the SST state in last ocean history record.
2). The SST may not be compacted to grib format correctly.